ProteinFormer: protein subcellular localization based on bioimages and modified pre-trained transformer

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Traditional protein localization experiments are costly and inefficient.
• Sequence-based methods struggle with protein translocation dynamics.
• Existing deep learning models lack global image integration for localization.

Purpose of the Study:

• To develop a novel deep learning model for protein subcellular localization.
• To integrate biological images with enhanced transformer architecture.
• To address challenges in small-sample scenarios and data scarcity.

Main Methods:

• Proposed ProteinFormer model combining ResNet for local features and a modified transformer for global fusion.
• Developed GL-ProteinFormer variant with residual learning, inductive bias, and ConvFFN for data scarcity.
• Utilized biological images for training and evaluation.

Main Results:

• ProteinFormer achieved state-of-the-art performance on Cyto_2017 dataset (91% single-label, 81% multi-label).
• GL-ProteinFormer showed superior generalization on the limited-sample IHC_2021 dataset (81%).
• ConvFFN improved accuracy by 4% and reduced computational costs.

Conclusions:

• ProteinFormer and GL-ProteinFormer outperform existing convolution-based methods.
• The fusion of biological images with transformer-based global modeling provides a robust solution.
• The approach is particularly effective for protein subcellular localization in data-limited settings.